DimitarV/eurospeech-bg-single-speaker

EuroSpeech BG — single-speaker subset

Bulgarian parliamentary speech from disco-eth/EuroSpeech, filtered down to clips containing exactly one speaker.

Why

EuroSpeech ships no speaker labels — its only identity-like field, video_id, is a parliamentary session containing dozens of speakers. To build LibriSpeechMix-style simulated mixtures for speaker-diarization training you first need clean single-speaker source audio. This subset is that source.

How clips were selected

Each segment is resampled to 16 kHz mono and passed through nvidia/diar_streaming_sortformer_4spk-v2. A clip is kept only if the diarizer reports exactly one speaker, where a second speaker counts only if it holds ≥25% of the primary speaker's speaking time (a threshold set because spurious secondary speakers were consistently <15% while genuine pairs were 41–94% balanced).

Selection deliberately favours precision over recall: ~2,200 h are available, so discarding good clips is cheap, whereas admitting a two-speaker clip would silently corrupt every mixture built from it.

Cheaper embedding heuristics (max-spread from the clip mean, spectral eigengap, adjacent-window jump counts) were evaluated first and all rejected — each assumed a speaker count or a single change point, and so scored a 4-speaker recording as more single-speaker than a 2-speaker one.

Fields

field	meaning
audio	16 kHz mono PCM16
text	official stenographic transcript (editorially cleaned, not verbatim)
asr_transcript	whisper-large-v3-turbo output, from EuroSpeech
cer	character error rate between the two, from EuroSpeech
video_id	parliamentary session — not a speaker id
sf_speaker_seconds	per-speaker speaking time reported by Sortformer

Caveats

• Speaker identities are inferred, not verified. See speaker_labels_v1.json. They are clustering output, not ground truth: a label means "these clips share a voice", never a named person. No official speaker annotation exists for this corpus.
• Transcripts are not verbatim. Stenographic records are editorially cleaned; disfluencies and false starts are typically removed. cer against the ASR transcript is a usable proxy for how heavily a segment was edited.
• Filter precision is not exhaustively validated. It was checked against a small calibration set and by listening to samples of both decisions.

Attribution

Derived from EuroSpeech (disco-eth), itself built from Bulgarian National Assembly proceedings. Speaker-count annotations are derived via Sortformer, not official.

Speaker labels (v1)

speaker_labels_v1.json maps clip id -> integer speaker id; speaker_summary_v1.json records the method and its parameters.

labels = json.load(open("speaker_labels_v1.json"))
speaker = labels[row["id"]]

clips labelled	502,671
sessions	1,722
session-level speakers	31,907
global speakers	4,660
speakers in >=2 sessions	1,800
thresholds	within 0.45, across 0.35

Derived by clustering TitaNet-L embeddings in two stages: within each video_id, then across session-level centroids, so a speaker's clips are reachable from multiple sittings. The distribution is heavy-tailed — the largest speaker holds 12,315 clips across hundreds of sittings (a presiding-officer profile), while 2,637 speakers hold a single clip.

Do not treat these as verified identity. They were checked by sweeping the clustering threshold and by listening to paired sub-clusters of the largest speakers, not by exhaustive validation. Labels are versioned because retuning changes every id.